1. No category

Why does the US dominate university league tables?
Mei Li, Sriram Shankar, and Kam Ki Tang , School of Economics Discussion Paper No. 391, June 2009, School of
Economics, The University of Queensland. Australia.
Full text available as:
PDF- Requires Adobe Acrobat Reader or other PDF viewer
ABSTRACT:
According to Academic Ranking of World Universities, the world’s top 500 universities are owned by only 38
countries, with the US alone having 157 of them. This paper investigates the socioeconomic determinants of the
wide performance gap between countries and whether the US’s dominance in the league table is largely due to
its economic power or something else. It is found that a large amount of cross country variation in university
performance can be explained by just four socioeconomic factors: income, population size, R&D spending, and
the national language. It is also found that conditional on the resources that it has, the US is actually
underperforming by about 4 to 10 percent.
EPrint Type:
Keywords:
Departmental Technical Report
Poisson model, Tobit model, count data, higher education, ranking
Subjects:
ID Code:
Deposited By:
JEL Classification: C2, I2
Why does the US dominate university league tables?
Mei Li, Sriram Shankar, and Kam Ki Tang*, #
The University of Queensland, School of Economics, QLD 4072, Australia
Abstract
According to Academic Ranking of World Universities, the world’s top 500
universities are owned by only 38 countries, with the US alone having 157 of them. This
paper investigates the socioeconomic determinants of the wide performance gap between
countries and whether the US’s dominance in the league table is largely due to its economic
power or something else. It is found that a large amount of cross country variation in
university performance can be explained by just four socioeconomic factors: income,
population size, R&D spending, and the national language. It is also found that conditional on
the resources that it has, the US is actually underperforming by about 4 to 10 percent.
JEL Classification: C2, I2
Keywords: Poisson model, Tobit model, count data, higher education, ranking
* Corresponding author. Address: School of Economics, University of Queensland, QLD
4072 Australia. Email: [email protected]. Fax: +617 33657299.
#
We would like to thank Chris O’Donnell, James Laurenceson and Alan Duhs for their
useful comments on earlier drafts. All remaining errors are obviously ours.
1
1. Introduction
In 2003, Shanghai Jiaotong University published the first Academic Ranking of
World Universities (ARWU). Another league table, the THES-QS World University
Rankings, was published a year later by Times Higher Education and Quacquarelli Symonds,
followed by many other similar endeavours subsequently.1 Not surprisingly, American
universities dominate all these world university league tables. Why the US ‘monopolizes’ the
world’s top universities is an intriguing question.
The objective of this study is to identify the key socioeconomic factors that determine
countries’ performance in world university league tables. In doing so, we aim to shed light on
the question posed in the title of the paper. The answer to this question has important policy
implications. If the US’s dominance in the league tables is more than just a result of its
economic hegemony, then there is a great need to understand the factors at the micro or
institutional level that are driving the success of its universities as it would provide valuable
learning opportunities to other lagging countries in improving their university education
systems. For example, Aghion et al. (2009) analyse how university governance affects
research output at European and U.S. universities, after realising U.S. universities are obvious
positive outliers in ranking performances. Also, in 2008, the University of Melbourne
restructured its academic programs entirely based on a hybrid of the US university and
European Bologna models, while all other Australian universities are watching as bystanders
with both interest and reservation.
The ranking on league tables increasingly have real resource implications for
universities. This is because, despite the criticisms of their accuracy, reliability and
usefulness, university rankings have been quickly adopted as a quality assurance mechanism
around the world. In particular, international students use league tables to help identify top
1
The Melbourne Institute has been ranking Australian universities (Williams & Dyke 2004; Williams 2007),
and U.S. News & World Report published its first World’s Best Colleges and Universities rankings in 2008
using the THE-QS data, after publishing America’s Best Colleges and America’s Best Graduate Schools for 25
years.
2
universities (Hazelkorn 2008). Research commissioned by the Higher Education Funding
Council of England indicates that league tables are already influencing international
recruitment, although more so in some markets than in others (HEFCE 2008). In Qatar,
scholarships for study abroad are limited to students going to highly ranked institutions as
identified by the ARWU and THES-QS rankings (Salmi & Saroyan 2007). Given this, it is
not surprising that many institutions have already used their ranking positions for marketing
purposes. Some institutions have even established a formal internal mechanism for reviewing
their rank and the majority of them have initiated actions in order to improve their
international ranking (Hazelkorn 2007; Healy 2009).
The importance of university league tables goes beyond the interests of prospective
students and university administrators. Education has become increasingly an export item.
For example, for Australia, which is one of the major players in international education,
education constitutes the third largest export item, contributing US$10,230 million to its total
exports in the year 2007-08 (DFAT 2008). Furthermore, as universities are a key place for
conducting frontier research and a place for higher education, poor performance in league
tables may indicate constraints on future development and growth.2
Due to the aforementioned factors, university rankings seem to be influencing policy
makers and possibly the classification of institutions and the allocation of funding (Hazelkorn
2007). Authorities in many countries have already responded to the international competition
in higher education. In 2005, Malaysia’s opposition called for a Royal commission of inquiry
in response to a fall of a hundred places of its top two universities in the THES-QS ranking
(Salmi & Saroyan 2007).3 In April 2008, the French Minister for Higher Education and
Research outlined that one of her priorities is to reinforce quality assurance of higher
2
A large body of literature has found important positive effects of education on economic growth and
development, e.g. Barro (1991), Schofer et al. (2000) and Koop et al (2000).
3
See section 3 for a critic of the volatile nature of THES-QS rankings.
3
education, which will entail a thorough analysis of international quality indicators, as well as
the impact of global rankings.
A common feature across different league tables is that the distribution of the world’s
top universities is highly skewed toward a single country—the US. According to the most
widely cited league table, ARWU, in 2008 the US has 17 top 20 universities identified in the
table, 54 top 100, 90 top 200, 114 top 300, 139 top 400, and 159 top 500.4 In comparison,
China, though having a much larger higher education sector5, has only 18 top 500
universities, 7 top 400, 6 top 300, but none in the top 200. Since university rankings in these
league tables are determined by the statistical indicators underpinning the ranking score, such
as the number of journal publications and the number of Nobel laureates, the ‘reason’ that the
US is well ahead of other countries in the league tables is simply that its universities have
published many more journal articles, recruited many more Nobel laureates and so forth.
Simply put, the league tables indicate that the US has more of a specific type of human
capital—academic talent—than other countries. Therefore, in trying to identify the key
socioeconomic factors that determine countries’ performance in university league tables, we
are essentially seeking for the socioeconomic explanation of cross country differences in
academic talent stocks.
To the best of our knowledge, there is no such quantitative analysis in the literature.
Aghion et al (2009) use the ARWU league tables to measure university output in their
examination of the effect of university governance. The focus of their analysis is on the
institutional level, not on the macroeconomic level. Also, they focus on the comparison
between the performances of European and US universities only. Marginson (2007) attempts
to look at this issue; however, he only compares countries’ share of top universities with their
4
Different league tables indicate the dominance of US universities to a different degree. For instance, the
THES-QS ranking noticeably puts more UK and Australian universities in top spots than the ARWU does.
However, the overall picture is largely the same.
5
According to the UNESCO, in 2006, 5.6 million tertiary students graduated in China, as compared to 2.6
million in the US.
4
share of economic capacities measured by total GDP multiplied by GDP per capita. By using
such simple summary statistics, the study does not control for other factors or noise in the
data. More importantly, by considering only those countries that successfully enter the league
tables, it confronts selection bias problems. In the current paper, we use regression methods
to address these problems.
This paper is inspired by a number of quantitative analyses of Olympic medal tally
results, including Bernard & Busse (2004), Johnson & Ali (2004) and Morton (2002).
However, our modelling technique is different in the way that we employ count data models
instead of the standard OLS regression or the Tobit model used by these studies. We
demonstrate that the Poisson model provides a better modelling approach in the current
context. Furthermore, as against the Olympic Games literature, endogeneity is potentially a
much bigger problem in the current study. This is because, to the extent that a country’s
performance in university league tables is a proxy for its research capability, the causality
from rankings to income cannot be ignored. Endogeneity is tested as part of the empirical
analysis.
The rest of the paper is organized as follows. Section 2 outlines the methodology,
while Section 3 explains the dataset used for the empirical work. Section 4 reports and
discusses the findings, and Section 5 concludes.
2. The Methodology
Academic talent can be considered a specific type of human capital, in contrast to
other types of human capital such as athletic talent or entrepreneurship. Similar to all other
types of human capital, academic talent cannot be directly measured. What can be measured
is the academic performance of people. A university’s research and publication performance,
and thus, ranking, reflects the size and quality of its stock of academic talent. Accordingly, a
country’s performance in the university ranking tally is an indication of the country’s stock of
5
this specific type of human capital. This study borrows from, and extends the methodology
used in a niche Olympic Games literature, particularly Bernard and Busse (2004). This
literature examines the socioeconomic determinants of countries’ performance in the
Olympic Games medal tally; likewise, here we seek to identify the key socioeconomic
determinants of countries’ performance in the university ranking tally.
In this paper, we focus on the ARWU league tables for reasons to be explained in the
data section. The ARWU dataset provides the ranking of up to 500 universities in the world.
This allows us to compute the number of top 500 universities that a country has in a
particular year; we denote this variable by TOP500i , where i indicates the country index.
Later on, we also consider smaller subsets like top 300 and top 100 universities. The value of
these variables for the vast majority of countries varies little over the six year period over
which the ARWU league tables have been compiled: 2003-2008. The only notable exception
is China whose share in the top 500 universities doubled during this period. The dataset is
essentially cross-sectional rather than panel in nature. This does not necessarily cause any
concerns because our primary interest is in identifying factors that explain the performance
gap between countries, and therefore cross country analysis will suffice. The empirical work
will focus only on the data for the latest year, 2008. We use the value of a single year rather
than the sample average because, as explained below, in Poisson regressions the dependent
variable must be an integer.
A large number of countries do not have a single university breaking into the ranks of
the top 500. However, these countries should also be included in the analysis to avoid
selection bias. To accommodate for zero and non-negative observations, we consider two
basic modelling options: the Tobit model and the Poisson model.
The Tobit model is typically considered as a solution to modelling non-negative data
samples containing a high proportion of zero observations. To motivate the use of the Tobit
6
model, we follow Bernard and Busse (2004) and model TOP500i as a function of the size of
the country’s academic talent stock, Ti , over some threshold level, T * , which in turn is a
function of the world average level of academic talent stock. 6 Since academic talent is not
directly observable, it is a latent variable. The Tobit model is given by:
⎧ln (Ti / T * ) if ln (Ti / T * ) > 0
⎪
yi = ⎨
0
if ln (Ti / T * ) ≤ 0
⎪⎩
(1)
where yi is the observed value of TOP500i for a given year.
To investigate what socioeconomic factors affect the accumulation of academic talent,
it is further hypothesized that
ln(Ti / T * ) = f (x′i β) + ei , ei ~ iid (0, σ 2 )
(2)
where xi is a vector of independent, socioeconomic factors and β the corresponding
unknown coefficients. Since the dependent variable is a measure of a country’s talent stock
versus the world average, the dependent variables should also be measured relative to their
world average values. Nevertheless, in the case of cross sectional analysis, the world average
values of the independent variables can simply be absorbed into the constant term and
thereby do not need to be accounted for explicitly.
A merit of using the Tobit model is that it captures the non-negative nature of the data
and confines the predicted values of the dependent variable to be non-negative, i.e. yˆi ≥ 0 .
This is, however, only one of the two restrictions, the other restriction being
6
∑
i
yˆi = 500 . In
Bernard and Busse’s specification of M it* in the Tobit model (equations 3 and 4 in their paper) is incorrect
(
)
because by setting M it* = ln Tit / ∑ j T ji , a log function of world medal share, M it* must always be nonpositive. Nevertheless, this does not affect their empirical analysis as one could alternatively set
M it* = ln(Tit / T * ) , where T * represents some threshold level of athletic talent. In fact, this is exactly the
specification used by Bernard and Busse in their working paper (2000). The use of panel data in their case,
however, raises another issue in that T * could be time varying because if the Olympic Games becomes more
competitive overtime, the athletic talent threshold to earn a medal will also increase. This is not an issue in the
current paper because we are using a cross sectional dataset.
7
OLS regressions, this restriction will be automatically satisfied because the expected value of
the error term is assumed to be zero. However, in the case of Tobit regressions, even though
the mean value of ei in equation (2) is assumed to be zero, the forecast errors of the zero
observations are not constrained to sum to zero.7 As a result, the second restriction will not be
satisfied, except by coincidence. While it is possible to rescale the predicted values so that the
predicted sum is the same as the actual sum, any over (under) prediction actually indicates
that overall, the model overestimates (underestimates) the coefficients. Unless by mere
coincidence the degrees of overestimation or underestimation for different coefficients are the
same, one cannot recover the ‘true’ coefficient values by rescaling the coefficient estimates
with the same factor.
A potential solution to the above problem is to estimate the Tobit model using
Bayesian estimation methods. Using Bayesian methods, we can make prior assumptions
about the properties of the dependent variable and restrict the total sum of the predicted value
to lie within a certain band around the actual sum. However, restricting the total sum to a
narrow band would require simulation of an extremely large Markov chain. Furthermore, the
Tobit model assumes that the latent variable has a continuous distribution. However, when
the concerned variable is truly discrete, standard methods like maximum likelihood Tobit
result in inconsistent parameter estimates (Mullahy 1986). Portney and Mullahy (1986) in
their application conduct Tobit specification error tests of Nelson (1981) and Lin and
Schmidt (1984) for their count measure and find considerable evidence of misspecification.
Rosenzweig and Wolpin (1982) in their study on the impact of governmental efforts to reduce
7
The log likelihood for censored regression is
( y − x ' β )2
x 'β
−1
ln L =
[log(2π ) + log(σ 2 ) + i 2i
]+
ln[1 − Φ ( i 2 )] .
2
σ
σ
y >0
y =0
∑
i
∑
i
The two parts correspond to the classical regression for the nonlimit observations and the relevant probabilities
for the limit observation, respectively. While the first order condition for the constant term forces the sum of the
residual term to be zero in the case of classical linear regressions, this is not the case for the Tobit model due to
the presence of the discrete term in the above log likelihood function.
8
population growth and augment child survival and schooling in a developing country find
that for their discrete fertility variable the estimated coefficient standard errors from the
fertility equation to be biased. Hence, they unambiguously rule out the use of the Tobit model
in their application.
An alternative to the Tobit model is the Poisson model. The Poisson distribution can
be written as
Pr (TOP500i = yi ) =
exp(−λi )λi yi
, λi ∈
yi !
+
, yi = 0,1, 2,...
(3)
where
E (TOP500i ) = var(Yi ) = λi
(equidispersion)
(4)
To ensure λi to be non-negative, the most common specification for λ is an
exponential function
λi = exp(x′i β)
(5)
where x is a vector of independent variables and β the corresponding unknown coefficients.
Equations (4) and (5) together establish the Poisson regression model
E (TOP500i | xi ) = λi = exp(x′i β)
(6)
The Poisson regression model has a number of merits: it captures the discrete and nonnegative nature of the data; it allows inference to be drawn on the probability of the event
occurrence; it allows for straightforward treatment of zero observations; it naturally accounts
for the heterokedastic and skewed distribution inherent to non-negative data (Winkelmann
and Zimmermann 1995). In the current context, it has an additional merit that it guarantees
the sum of the predicted values to be equal to the actual sum. Assuming independence among
the count variables, the log-likelihood for the Poisson regression model follows as:
N
L = ∑ yi xi′β − exp(xi′β) − ln( yi !)
i =1
9
(7)
The first order conditions, ∂L / ∂ β = 0 , yield a system of equations (one for each β) of the
form:
N
∑ ( y − exp(x ′β))x
i =1
i
i
i
=0
(8)
Equation (8) ensures that the sum of predicted values is the same as the actual sum if x
contains a constant term.
A potential problem confronting the Poisson model is that in many applications, the
assumption of “equidispersion” is violated. For instance, in our dataset the mean of TOP500i
is 5.33 and standard derivation 17.99 (see Table 2). This is referred as “overdispersion” in the
literature,8 and it could be eliminated to a certain extent by the inclusion of additional
regressors. Alternatively, one can specify a distribution that permits more flexible modelling
of the variance than the Poisson distribution. The standard alternative distribution used is
negative binomial (NEGBIN), with variance either assumed to be a linear or a quadratic
function of the mean (Cameron and Trivedi 1986). A much simpler alternative is to use the
Generalized Linear Model (GLM) under the assumption that the true variance is proportional
to the distribution used to specify the log likelihood:
var(TOP500i | xi ) = σ 2 varML (TOP500i | xi )
(9)
If there is overdispersion, one would expect σ 2 > 1 .
The conventional R-squared statistics based on a sum of squared residuals does not
apply to either the Poisson or the Tobit model. Wooldridge (2009) suggests calculating the R2
squared as the squared correlation coefficient between yi and yˆi , i.e. R% 2 = ( corr ( yi , yˆi ) ) ,
motivated by the fact that the usual R-squared for OLS regressions is equal to the squared
correlation between yi and the OLS fitted values.
8
However, as it is shown later our sample actually passes the overdispersion test.
10
Beside correlation, another, yet more direct, measure of goodness-of-fit is simply the
extent to which the model can predict the actual value of the dependent variable for given
values of the independent variables. Since the Poisson model already guarantees the sum of
the predicted values to be equal to the actual sum, i.e.
∑ yˆ = ∑ y , we can use the sum of
i
absolute forecast error as a ratio of the total actual value, i.e.
i
∑ y − yˆ / ∑ y
i
i
i
as a measure
of goodness-of-fit. Yet, this measure is bounded from below at 0 but unbounded from above.
In order to make the measure lie between 0 and 1 we propose to construct
R 2 = ∑ yi / ( ∑ yi + yi − yˆi
) as an alternative indicator of the goodness-of-fit. The value of
R 2 increases as forecast errors reduce, with 1 indicating a perfect match for every single
observation and 0 indicating at least one infinitely large forecast error. A value of R 2 equal
to, say, 0.8 means that the average forecast error is approximately equal to 20 percent of the
actual value. One should, however, remember that the Poisson and Tobit estimates are chosen
to maximize the log-likelihood functions, not R-squared. As a result, unlike OLS regressions,
adding more variables into the model does not necessarily improve R% 2 or R 2 .
3. Data
There are numerous organizations publishing international university rankings,
including Shanghai Jiaotong University’s Academic Ranking of World Universities
(ARWU) 9 , the Times Higher-Quacquarelli Symonds World University Rankings (THESQS)10, the Webometrics Ranking of World Universities11 and Newsweek magazine’s Top 500
Global Universities 12 . Amongst them the ARWU and THES-QS rankings are most wellknown and widely cited by academics and the media, e.g. The Economist (2005). A
consensus is emerging in the literature that the ARWU, despite its limitations, provides a
9
http://www.arwu.org
http://www.topuniversities.com
11
http://www.webometrics.info/about.html
12
http://web.archive.org/web/20060820193615/http://msnbc.msn.com/id/14321230/site/newsweek
11
10
superior indicator of university excellence in terms of objectivity and comprehensiveness
(Marginson, 2007; Taylor & Braddock, 2007; Hazelkorn, 2007; Buela-Casal et al. 2007). The
ARWU aims to measure the institutions’ research strength using internationally comparable
indicators such as number of alumni and staff winning Nobel Prizes and Fields Medals,
publication count, as well as a per capita performance measure (Liu & Cheng 2005). On the
contrary, the THES-QS ranking relies heavily on peer reviews, which are criticized for being
strongly subjective and resulting in the high volatility of the rankings. Therefore, in this paper
we focus only on the ranking statistics of ARWU.
For the independent variables, we examine a number of socioeconomic variables that
potentially affect the accumulation of academic talent. Our benchmark model includes four
independent variables: log population size ( LPOP ), log income ( LGDPPC ), R&D spending
as a percentage of GDP ( R _ D ), and a dummy for English as the native language ( ENG ). If
academic talent is randomly distributed around the world, then other things being equal,
countries with a larger population should have a larger academic talent stock. However,
academic talent, like athletic talent, can be a result of nurture as well as of nature; therefore,
the amount of available resources matters. We use income as a general measure of the
financial resources available in a country, and use total expenditure on R&D as a specific
measure because if the total R&D expenditure of a country goes up, the amount going to the
higher education sector is likely to go up accordingly. The English language dummy is to test
if English speaking countries are in a privileged position. At first glance, the answer seems to
be an obvious yes given that the rankings primarily consider publications in English.
However, many academic staff in non-English speaking countries now teach and publish in
English.13 Moreover, universities in English speaking countries have the advantage of being
13
Japanese universities, for instance, offer many postgraduate programs to “international or internationallyminded” students, and recruit international scholars without requiring the knowledge of Japanese language
(Hazelkorn 2008).
12
able to recruit both native and non-native English speaking academics from around the world,
whereas universities in non-English speaking countries are more confined in their recruitment
if their staff members are required to teach in the local language. In summary, positive signs
are expected for all these four variables.
Besides these four variables, we also try to include a number of other socioeconomic
variables for robustness tests. These include the number of migrants as a percentage of total
population (MIG), public expenditure on education as a percentage of GDP (EDUGDP), log
public expenditure per tertiary student (LEXPTER), and the number of tertiary students as a
percentage of population (TERPOP). The migration variable is to account for the fact that
academics are highly mobile internationally. Public expenditure on education is yet another
possible measure of resources. Different from the R&D expenditure, EDUGDP captures the
resources flowing into all primary, secondary and tertiary sectors. The deployment of this
variable can be motivated by the fact that, if a country’s primary and secondary education is
sound, it can produce good prospective tertiary students and academics. In comparison, the
expenditure per tertiary student specifically measures the resources put into the higher
education sector. Lastly, the number of tertiary students as a proportion of the population size
measures the size of the higher education sector; countries with a larger higher education
sector obviously need to employ more academics. Therefore, all four additional variables are
expected to have positive signs on their respective coefficients. These additional independent
variables, however, are found to be statistically insignificant and are subsequently dropped.
In the next section, we report and discuss only the results of the key variables.14
Table 1 lists all the variables used in this paper and their sources and Table 2 shows
their summary statistics. To mitigate reverse causality, the time periods of the independent
14
The results of models with additional variables can be obtained from authors on request.
13
variables are chosen to lag 2008 (we formally test for reverse causality later on). The data for
individual countries are provided in Table A1.
A few observations on the figures in Table 2 are worth commenting. First, the total
sum of TOP500i is slightly less than 500 because some countries/regions that have top 500
universities, like Taiwan, are excluded due to the lack of data for other variables. Second,
besides TOP500i , the table also shows the statistics of TOP300i and TOP100i . In the case of
TOP300i , the total sum is larger than 300 because some ranking positions are shared by more
than one university. Third, there is a high degree of concentration of top universities amongst
a small number of countries. For instance, only 40 percent of the sample countries (i.e. 38 out
of 93) have one or more top 500 universities, and the percentage falls to 15 percent as it
comes to the top 100 universities. In other words, over half the countries in the sample have a
zero value for their dependent variables. Lastly, when the additional independent variables
are added, the sample size reduces from 93 to 79.
4 Empirical Results
4.1 Results for TOP500
We first report the Poisson model results for TOP500i . The results for various model
specifications are shown in Table 3. Models (a) and (b) show that either income or population
size by itself does a very poor job in explaining TOP500i ; but once the two are combined as
a single measure of the total economic size in model (c), the explanatory power improves
dramatically. Model (d) further shows that LGDPPC has twice the effect of LPOP, indicating
that income has individual effects even after controlling for the total economic size. The
variables have the expected signs in all three models. Model (e) shows that adding R _ D into
the model only marginally improves its goodness-of-fit, and the variable is just significant at
the 10 percent level with the expected sign. Model (f) further adds ENG as another control
14
variable. It shows that the English language dummy has the expected sign and is significant at
a level slightly higher than 5 percent. The inclusion of ENG also improves the significance of
R _ D to the 1 percent level. Despite both ENG and R _ D being highly significant
individually, their inclusion does not add much to the overall explanatory power of the model
as compared to the more parsimonious model (d). In what follows, we consider model (f) as
the benchmark model.
According to the benchmark model (f), other factors being equal, expanding the
population size by one percent increases the number of top 500 universities in a country by
0.77 percent, while a one-percent rise in the income level increases that figure by nearly 1.5
percent. The model also shows that, other factors being equal, raising the expenditure of
R&D by one percentage point of GDP increases the number of top 500 universities by 28
percent. While the effect may appear to be very large, a one percentage point increase in
R&D actually represents more than doubling the R&D spending for an average country in the
sample (the mean value of R&D is equal to 0.89 percent of GDP). Lastly, it is shown that an
English-speaking country has 41 percent15 more top 500 universities than a non-English
speaking, but otherwise identical country.
Figure 1 shows the actual values and the prediction errors (actual minus predicted
value) of TOP500i for each of the 38 countries that have at least one top 500 university. The
full results are shown in Table 4, which reports the predicted values of TOP500i for each of
the 93 countries in the sample, the prediction errors (the actual minus the predicted value) and
the prediction errors in proportional terms if the actual value is bigger than zero. Although the
US is the frontrunner in the ARWU league table, it actually underperforms by nearly 18
universities based on the benchmark model, followed by Japan (17 universities short).
However, in proportional terms, Mexico is the most underperformed country with a shortfall
15
Since ENG is a dummy variable, the effect of a change of its value from 0 to 1 is computed as
exp(0.344*1)/exp(0.344*0) – 1 = 0.41.
15
of 332 percent, followed by Russia with a shortfall of 186 percent. For countries that have
zero actual observations, Iran is predicted to have the largest number of top 500 universities
of about 2, followed by Malaysia with a predicted number of 1.5 and Kuwait of 1.4. The
results indicate that, given these countries’ human and non-human resources, they have the
potential to perform better. This in turn suggests that there are some factors hindering them
from fully realizing their potential, such as insufficient resources being channelled into the
higher education sector (e.g. Mexico16), institutional barriers to participating in international
academic communities (e.g. Iran17), or salary incentives being unconducive to publishing in
high ranked international journals (e.g. Japan).
On the other hand, the UK outperforms the model prediction by 15 universities, or 36
percent in percentage terms. The next best ‘outperformer’ is a rising economic and political
power—China, which has 13 universities, or 70 percent, more than the model’s prediction.
For countries with non-zero top 500 universities, China has the best outperformance in
proportional terms. The third best outperformer in absolute terms is Germany, which has
about 10 more universities than the model’s prediction.18
Despite the aforementioned theoretical limitations of the Tobit model, it is useful to
examine their empirical significance. The results for the Tobit model based on the benchmark
model are also reported in Table 5 along with the Poisson model results for the ease of
comparison. It can be seen that the signs for all variables remain correct in the Tobit
regression. In terms of goodness-of-fit, both R% 2 and R 2 statistics indicate that the Poisson
model provides a better fit for the data than the Tobit model. In fact, the total predicted
16
The higher education system in Mexico has undergone a long process of transition, which is regarded as “not
necessarily positive” , and a range of problems have been identified, such as insufficient enrolment levels, poor
staff salary, demoralized faculty, and non-existing student financial aid (Ordorika 1996).
17
There are unverified claims that Iranian scholars and universities sometimes experience formal and informal
barriers to international collaboration, such as being denied visa to attend academic conferences overseas
(Pankratz 2008).
18
Interestingly, the German government, after acknowledging the country has lost the glory of being one of the
world’s intellectual centres, has put forward a policy of creating a German Ivy League (Vogel 2006).
16
number of TOP500i by the Tobit model is equal to 663.4, which is 33 percent more than the
actual sum of 496! More importantly, the fact that the Tobit model over predicts TOP500i
implies that overall it overestimates the coefficients of the independent variables. An
explanation for this result is that there is a high level of heteroskedasticity in the data set, as
commonly seen in cross sectional data. In OLS regressions, heteroskedasticity affects the
standard errors, but the estimates for the coefficients remain unbiased. However, in the case
of the Tobit model, this will lead to biased estimates of the coefficients.19 The Poisson model,
on the contrary, automatically accounts for heteroskedasticity (see equation (5)).
4.2 Robustness Tests
The US dominates the ARWU league table with a total of 159 top 500 universities,
equivalent to 31.8 percent. The UK is the next most prominent country on the league table,
with a total of 42 top 500 universities, equivalent to 8.4 percent. Together they account for 40
percent of the top 500 universities. Therefore, it is important to test the robustness of the
results with respect to the exclusion of these two ‘outliers’. The results are also shown in
Table 5. The exclusion of the US and the UK has a very small effect on the coefficient
values and virtually no impacts on their significance or signs for most variables. The only
exception is that the significance of the English language dummy becomes significant at the l
percent level and the magnitude of its coefficient also increases. Therefore, we can conclude
that, the model for TOP500i is very robust to the exclusion of the top two performers.
A key distinction between the model for the university ranking and that for the
Olympics Games is that the chance of having reverse causality onto income is higher for the
former than for the latter. This is because academic talent is more likely to have bigger
impacts on the macroeconomy than athletic talent due to the spillover effects of knowledge
creation. Besides income, the expenditure on R&D may also incur endogeneity problems in
19
See David Madigan’s note on Logistic and Tobit Regression, available at
http://stat.rutgers.edu/~madigan/COLUMBIA/
17
that top universities are more attractive to research funding. In testing endogeniety in the
Poisson model we follow the procedure described by Wooldridge (2002), which requires
instrumental variables for the potentially endogenous variables. One obvious choice of
instrument for a potentially endogenous variable is its lagged value. Due to data limitation,
we use the 1980 values of LGDPPCi and the 1996 values of R _ Di as their respective
instruments. We first estimate the following reduced form OLS regressions:
LGDPPCi = c1 + c2 LGDPPC 1980, i + ui
(10)
R _ Di = c3 + c4 R _ D1996, i + vi
(11)
and
Once we obtain the residuals uˆi and vˆi , we estimate the Poisson regression model by
regressing the count data on the explanatory variables and the residuals. The Wald test that
the coefficients of the residuals are zero, cannot be rejected at the 5 percent significance level,
indicating that there is insufficient evidence to support the claim that income and R&D
expenditure are endogenous in the benchmark model. There are a number of possible
explanations for the absence of evidence of reverse causality from TOP500i to income.
Firstly, TOP500i only captures a specific type of human capital within the total human
capital stock of a country; secondly, the research output used in determining the ARWU
rankings is mostly academic research instead of commercial research; thirdly, the
materialization of academic research in terms of commercialization or policy making may
take years; and lastly, the benefit of academic research outcome often spills over to other
countries, diluting the effect of difference in league table performance on income difference.
Likewise, the lack of evidence of reserve causality from TOP500i to R&D spending may be
due to the fact that the effect of TOP500i on university R&D spending may be too small to
be discernible in the much bigger pool of national R&D spending.
18
Lastly, recall that when there is overdispersion, a negative binomial (NEGBIN)
distribution is more appropriate than the Poisson distribution. We employ the procedure
suggested by Wooldridge (2002) to test for overdispersion and to estimate the overdispersion
parameter α for the Poisson model. The test suggests that when there is an overdispersion the
variance is greater than the mean, i.e. var( yi ) = μi (1 + αμi ) , where μi = exp( X i β ) . The null
hypothesis ( α = 0 ) that there is no overdispersion cannot be rejected at the 5 percent
significance level, indicating that the Poisson model restriction that the mean is equal to the
variance holds well.
4.3 Results for TOP300 and TOP100
Besides top 500 universities, we also consider countries’ performance in more elite
subgroups, top 300 and top 100 universities. The dependent variables are denoted
respectively as TOP300i and TOP100i . These results are reported in Table 6 while the results
for TOP500i are also repeated in the table for the ease of comparison.
The signs of all variables remain the same across all three models. The goodness-offit of the three models is also very similar, indicating that the benchmark model maintains its
explanatory power for all three dependent variables. Nonetheless, there are gradual changes
of the magnitude of the coefficients. Specifically, the coefficient of population size reduces as
it moves toward the more elite group (from 0.77 for TOP500i to 0.67 for TOP100i ). On the
contrary, the coefficients of income, R&D expenditure, and English language all go up. The
coefficient of income increases by 50 percent (from 1.50 to 2.27), while those of R&D
expenditure and the English dummy increase by around three fold (from 0.28 to 0.81 for
R _ D and from 0.34 to 1.09 for ENG). It can be inferred from these results that, the more
elite the academic talent, the more important the nurturing factors and the less random their
distribution across countries. The results are somewhat expected in that, as the old saying
19
goes: “genius is ten percent inspiration and ninety percent perspiration”; at the same time
perspiration requires sufficient resources and a good working environment, and their
distributions across countries are far from being random. Moreover, there is a strong
tendency for the most talented academics to move to places where they can be best supported
and most easily find their peers. Together these factors indicate that there is an agglomeration
effect of academic talent clustering in countries with better resources and working
environment.
Figure 2 shows the actual values and the prediction errors of TOP100i for each of the
15 countries that have at least one top 100 universities. The US continues to underperform,
but the margin shrinks from the 10 percent in the case of TOP500i to 4.3 for the more elite
top one hundred group. The margin of over performance for the UK also shrinks from 57
percent to 17.6 percent. The results are probably related to the fact that resources are,
expectedly, not uniformly distributed across universities. Inequality in resources within the
top 100 universities is likely to be smaller than within the top 500 universities. Therefore,
aggregate measures of national resources will be a better proxy of the resources available to a
smaller and more homogenous group of universities than those available to a larger, more
diversified one.
For the most elite universities, the dominance of the US and the UK is even more
striking. The two countries have 114 and 33 top 300 universities (38 percent and 11 percent),
and 54 and 11 top 100 universities (54 percent and 11 percent) respectively. Once again we
examine how robust the results for the most elite group of universities are by excluding these
two top performing countries. Due to space limitations, we only report the results for
TOP100i . When the US is excluded from the sample, there is an increase in the coefficients
of all the independent variables; when the UK is also excluded, the coefficient of population
size returns to close to the value of the full sample. Also, when both countries are excluded,
20
the coefficient of the English language dummy drops to less than the original value in the full
sample, and its significance level also drops down to the 10 percent level. Therefore, we
conclude that while overall the model remains robust with respect to the exclusion of the US
and the UK, expectedly it is somewhat more sensitive for TOP100i than for TOP500i .
5. Discussions and Conclusions
This paper aims to seek a better understanding of the socioeconomic determinants of
countries’ performance in university league tables. It focuses on the most widely cited
university league table—Shanghai Jiaotong University’s Academic Ranking of World
Universities (ARWU). The ARWU league table, like other league tables, is dominated by a
single country—the US. Sixty percent of the countries in the sample, on the contrary, do not
have a single university successfully breaking into the top 500 rank. The analysis in the paper
helps shed light on the large performance gap between countries in the league tables.
The empirical methodology needs to satisfy two restrictions: firstly, the predicted
number of top universities for any country should be non-negative; secondly, the total
predicted number of top universities should be identical to the actual total number. Standard
Tobit regressions satisfy the first restriction but not the second, while standard OLS
regressions satisfy the second but not the first one. Poisson regressions satisfy both
restrictions and thereby are applied to identify the key socioeconomic determinants of the
number of top 500 universities that countries had in 2008.
We find that a large proportion of cross countries difference in the ARWU league
tables can be readily explained by a few variables, primarily population size and income, with
the addition of R&D expenditure and an English language dummy. The findings confirm that
resources, including R&D funding, are crucial in building an internationally competitive
higher education sector. The model is robust to a number of specifications tests and exclusion
21
of the top two performers, the US and the UK. It is also robust to the modelling of top 100 or
top 300 universities.
Despite the observation that the US monopolizes the world’s top universities, our
finding indicates that there is nothing extraordinary about its performance. According to our
model, the reason for the US’s dominance is due to its large population and economic size,
further enhanced by its large expenditure on R&D (2.7 percent of GDP compared to the
sample mean of 0.89 percent) and its predominant language being English. In fact, given the
resources it has, the US is underperforming by about 4 to 10 percent based on the model
prediction. This finding does not necessarily come as a surprise in that on a very broad sense,
the similarities of university systems across countries are probably bigger than their
differences. This finding may nevertheless make sobering reading for administrators and
policymakers who fix their eyes firmly on US universities in searching for a better model for
their institutions.
On the other hand, one needs to be cautious in inferring from the findings, the relative
merits of the US university model. Despite the fact that the US’s performance falls short of
the model prediction, we must emphasise that it is still the frontrunner by a large margin.
Furthermore, the finding only indicates that down to top 100 universities, there is no
discernible country specific effect in the US to be explained. It does not exclude the
possibility that what distinguishes the US system from the others lies in the very top end of
the league tables, say, the top 50 or top 20 universities, where the dominance of the US is
strongest. Additionally, many academics in other countries were educated in US universities
(and the other way round as well), and therefore their success could be attributed at least
partly to the US model.
Another inference one might be tempted to draw from the findings is that, the US’s
current hegemonic position in the league tables could be challenged if its economic power
22
weakens. This is particularly relevant in the context that the US’s economic hegemony is
increasingly being tested by fast growing economies in Asia and Latin America, especially
the BRIC (Brazil, Russia, India, and China) countries. However, such a scenario of declining
US dominance in the league tables, while possible, is not necessarily warranted. For a given
total amount of resources provided for the higher education sector, how the resources are
distributed amongst universities can also make a significant difference to the overall national
outcome. Therefore, one must be careful in making predictions of a country’s future
performance simply based on a linear projection of its macroeconomic indicators.
The distributional factor may also explain why China is the best outperformer in the
top 500 league table in proportional terms. Although China is only a lower-middle income
country, its sheer size allows it to mobilize more resources than its income level suggests, and
focus them on a small number of universities. According to recent estimates (Li 2004),
approximately US$2.2 billion was distributed to a selected number of universities during the
period between 1996 and 2000. One of the programs launched by the Chinese government,
Project 985, was indeed designed to meet the policy objective that “China must have a
number of first-rate universities of international advanced level”.
The strong effect of English language on countries’ performance suggests that
universities in non-English speaking countries need concerted efforts in disseminating their
research output in English medium in order to improve their rankings. Endeavour in this
direction has already been observed in Europe and increasingly in Latin America and Asia,
implying that the ‘natural’ advantage of English-speaking countries will be gradually eroded
over time.
We conclude this paper by briefly discussing its limitations and providing pointers for
future research opportunities. In focusing on socioeconomic factors, the analysis in this paper
23
is silent on the importance of institutional factors,20 such as student and staff numbers, the
size of endowment, incentive structure, and the status of a public/private university. The last
factor is of particular interest as it could have implications in policy debates, as to what extent
the higher education sector, which is largely a semi public sector in most countries, should be
allowed to be privatized. The importance of these factors may vary across different countries,
and how institutional factors may interact with national factors would also be worth
investigating. An analysis at the institutional level thereby will constitute a natural
progression of this line of research.
Reference
Aghion, P, Dewatripoint, M, Hoxby CM, Mas-Colell, A & Sapir A 2009 ‘The governance
and performance of research universities: evidence from Europe and the U.S.’ NEBR
Working Paper # 14851.
Barro, RJ 1991, 'Economic growth in a cross section of countries', The Quarterly Journal of
Economics, vol. 106, no. 2, pp. 407-43.
Bernard, AB & Busse, MR 2000, 'Who Wins the Olympic Games: Economic Development
and Medal Totals', National Bureau of Economic Research Working Paper Series,
vol. No. 7998.
---- 2004, 'Who Wins the Olympic Games: Economic Resources and Medal Totals', Review of
Economics and Statistics, vol. 86, no. 1, pp. 413-7.
Buela-Casal, G, Gutiérrez-Martínez, O, Bermúdez-Sánchez, M & Vadillo-Muñoz, O 2007,
'Comparative study of international academic rankings of universities',
Scientometrics, vol. 71, no. 3, pp. 349-65.
20
Aghion et al (2009) is a good example of research in this direction.
24
Cameron, AC & Trivedi, PK 1986, 'Econometric Models Based on Count Data: Comparisons
and Applications of Some Estimators and Tests', Journal of Applied Econometrics,
vol. 1, no. 1, pp. 29-53.
DFAT 2008, Composition of Trade Australia, Australian Government Department of Foreign
Affairs and Trade.
Hazelkorn, E 2007, 'The Impact of League Tables and Ranking Systems on Higher Education
Decision Making', Higher Education Management and Policy, vol. 19, no. 2.
---- 2008, From Globalisation to Internationalisation – the Influence of Rankings on
Japanese Higher Education, OECD.
---- 2008, 'Learning to Live with League Tables and Ranking: The Experience of Institutional
Leaders', High Education Policy, vol. 21, no. 2, pp. 193-215.
Healy, G 2009, 'Long way to the top for ranking role', The Australian, April 22.
Higher Education Funding Council for England 2008, Counting What is Measured or
Measuring What Counts? League Tables and Their Impact on Higher Education
Institutions in England, London.
Johnson, DKN & Ali, A 2004, 'A Tale of Two Seasons: Participation and Medal Counts at
the Summer and Winter Olympic Games', Social Science Quarterly, vol. 85, no. 4.
Koop, G, Osiewalski, J & Steel, MFJ 2000, 'Modeling the sources of output growth in a panel
of countries', Journal of Business & Economic Statistics, vol. 18, no. 3, pp. 284-99.
Li, L 2004, 'China's Higher Education Reform 1997-2003: A Summary', Asia Pacific
Education Review, vol. 1, pp. 14-22.
Lin, T-F & Schmidt, P 1984, 'A Test of the Tobit Specification Against an Alternative
Suggested by Cragg', The Review of Economics and Statistics, vol. 66, no. 1, pp. 1747.
25
Liu, NC & Cheng, Y 2005, 'Academic Ranking of World Universities–Methodologies and
Problems', Higher Education in Europe, vol. 30, no. 2.
Marginson, S 2007, 'Global University Rankings: Implications in General and for Australia',
Journal of Higher Education Policy and Management, vol. 29, no. 2, pp. 131-42.
Morton, RH 2002, 'Who won the Sydney 2000 Olympics?: an allometric approach', The
Statistician, vol. 51, pp. 147-55.
Mullahy, J 1986, 'Specification and testing of some modified count data models', Journal of
Econometrics, vol. 33, no. 3, pp. 341-65.
Nelson, FD 1981, 'A Test for Misspecification in the Censored Normal Model',
Econometrica, vol. 49, no. 5, pp. 1317-29.
Ordorika, I 1996, 'Mexican Higher Education in Transition: From Politically to Financially
Driven Public Policies', International Higher Education, no. 5.
Pankratz, J 2008, 'Iran is recognizing the value of higher education', The Record.
Portney, PR & Mullahy, J 1986, 'Urban air quality and acute respiratory illness', Journal of
Urban Economics, vol. 20, pp. 21-38.
Rosenzweig, MR & Wolpin, KI 1982, 'Governmental interventions and household behavior
in a developing country : Anticipating the unanticipated consequences of social
programs', Journal of Development Economics, vol. 10, no. 2, pp. 209-25.
Salmi, J & Saroyan, A 2007, 'League Tables as Policy Instruments: Uses and Misuses',
Higher Education Management and Policy, vol. 19, no. 2.
Schofer, E, Ramirez, FO & Meyer, JW 2000, 'The effects of science on national economic
development, 1970 to 1990', American Sociological Review, vol. 65, no. 6, pp. 86687.
26
Taylor, P & Braddock, R 2007, 'International University Ranking Systems and the Idea of
University Excellence', Journal of Higher Education Policy and Management, vol.
29, no. 3, pp. 245-60.
The Economist, 2005, ‘The brains business.” Sep 8th , the print edition.
Vogel, G 2006, 'A German Ivy League Takes Shape', Science, vol. 314.
Williams, R 2007, 'Ranking Australian Universities: Controlling for Scope', Melbourne
Institute, University of Melbourne.
Williams, R & Dyke, NV 2004, The International Standing of Australian Universities,
Melbourne Institute of Applied Economic and Social Research, Melbourne.
Winkelmann, R & Zimmermann, KF 1995, 'Recent Developments in Count Data Modelling:
Theory and Application', Journal of Economic Surveys, vol. 9, no. 1, pp. 1-24.
Wooldridge, JM 2002, Econometric analysis of cross section and panel data, Cambridge,
Mass.; London: MIT.
---- 2009, Introductory econometrics: a modern approach, 4th edn, South-Western Cengage
Learning, Mason, Ohio.
27
Table 1 Definitions and sources of variables.
Variable
Definition
Year
Source
TOP500
The number of top 500 universities
in ARWU.
Log total population
2008
ARWU
1999-2004
Average
1999-2004
Average
1999-2004
Average
WDI
+
The dependent variable; similar definitions for
TOP300 and TOP100
To measure the size effect
WDI
+
To measure the general resource effect
WDI
+
To measure the specific resource effect
CIA World
Factbook
+
To measure if being an English speaking country has
advantage
1999-2004
Average
WDI
+
1999-2004
Average
1999-2004
Average
WDI
+
To measure how open the labour market is (it is
assumed that academic labour market is of similar
degree of openness as the national labour market)
To measure the overall quality of the education system
WDI
+
To measure the quality of the higher education sector
1999-2004
Average
UNESCO
Institute for
Statistics
+
To measure the size of the higher education sector
LPOP
LGDPPC
R_D
ENG
MIG
EXPGDP
LEXPTER
TERPOP
Log GDP per capita, (PPP, constant
2000 international dollar)
Research and development
expenditure (% of GDP)
Dummy, = 1 if English is the first
language or widely used, = 0
otherwise
International migration stock (% of
population)
Public expenditure on education (%
of GDP)
Log public expenditure per tertiary
student (PPP, constant 2000
international dollar)
Number of tertiary student (% of
population)
28
Expected
sign
Note
Table 2 Summary statistics of the dependent and independent variables
TOP500
TOP300
TOP100
LPOP
LGDPPC
R_D
ENG
MIG
5.33
3.24
1.08
16.21
9.09
0.89
0.25
8.09
4.80
8.16
2.95
0
0
0
16.14
9.15
0.58
0
3.39
4.71
8.30
3.13
Maximum
159
114
54
20.97
11.07
4.42
1
64.11
11.03
10.82
6.34
Minimum
0
0
0
11.67
6.34
0.01
0
0.04
0.82
5.17
0.08
Std. Dev.
17.99
12.70
5.76
1.79
1.16
0.95
0.43
11.42
1.61
1.16
1.46
Skewness
7.03
7.46
8.54
-0.05
-0.47
1.58
1.17
2.91
0.60
-0.37
-0.14
Kurtosis
58.89
64.00
78.42
3.24
2.31
5.06
2.37
12.69
4.82
2.87
2.37
Sum
496
301
100
1507.23
844.92
82.69
23
639.17
378.82
645.00
232.85
No. of Obs.
93
93
93
93
93
93
93
79
79
79
79
No. of Obs. > 0
38
31
15
Mean
Median
29
EDUGDP LEXPTER TERPOP
Table 3 Estimation results of the Poisson model for TOP500
(a)
(b)
(c)
(d)
(e)
(f)
Constant
-8.818***
(2.475)
-13.809***
(4.637)
-27.254***
(1.823)
-31.658***
(1.766)
-29.297***
(2.125)
-26.677***
(2.072)
LPOP
0.611***
(0.150)
0.869***
(0.055)
0.826***
(0.048)
0.766***
(0.060)
1.877***
(0.230)
1.677***
(0.262)
1.499***
(0.231)
0.215*
(0.127)
0.279***
(0.098)
LGDPPC
1.580***
(0.474)
LPOP + LGDPPC
(=LGDP)
1.071***
(0.064)
R_D
ENG
0.344*
(0.180)
2
R
~
R2
No. obs.
0.430
0.469
0.635
0.767
0.768
0.768
0.096
0.109
0.892
0.966
0.957
0.965
93
93
93
93
93
93
Standard errors are reported in parentheses, and are Huber/White standard errors.
*, **, *** indicates significance at the 10%, 5%, and 1% level, respectively.
R2 =
∑ y / ( ∑ y + y − yˆ ) ; R%
i
i
i
i
2
= ( corr ( y, yˆ ) )
2
30
Table 4 The Actual and predicted number of top 500 universities
Argentina
Armenia
Australia
Austria
Azerbaijan
Belarus
Belgium
Bolivia
Brazil
Brunei Darussalam
Bulgaria
Cambodia
Canada
Chile
China
Colombia
Costa Rica
Croatia
Cyprus
Czech Republic
Denmark
Ecuador
Egypt, Arab Rep.
Estonia
Finland
France
Georgia
Germany
Greece
Hong Kong, China
Hungary
Iceland
India
Indonesia
Iran, Islamic Rep.
Ireland
Israel
Italy
Jamaica
Japan
Jordan
Kazakhstan
Korea, Rep.
Kuwait
Kyrgyz Republic
Latvia
Lesotho
Lithuania
Luxembourg
Macao, China
Actual
Top500
1
0
15
7
0
0
7
0
6
0
0
0
21
2
18
0
0
0
0
1
4
0
0
0
6
23
0
40
2
5
2
0
2
0
0
3
6
22
0
31
0
0
8
0
0
0
0
0
0
0
Predicted
Top500
1.74
0.04
11.15
5.23
0.09
0.36
5.47
0.12
4.88
0.47
0.37
0.03
20.41
1.10
5.35
0.75
0.24
0.52
0.30
1.94
4.09
0.32
0.75
0.22
4.21
21.21
0.05
29.89
2.96
3.78
1.33
0.46
2.79
0.93
2.15
3.45
4.61
13.87
0.13
47.64
0.09
0.44
9.87
1.41
0.02
0.21
0.01
0.35
1.38
0.23
31
Prediction
Error
-0.74
-0.04
3.85
1.77
-0.09
-0.36
1.53
-0.12
1.12
-0.47
-0.37
-0.03
0.59
0.90
12.65
-0.75
-0.24
-0.52
-0.30
-0.94
-0.09
-0.32
-0.75
-0.22
1.79
1.79
-0.05
10.11
-0.96
1.22
0.67
-0.46
-0.79
-0.93
-2.15
-0.45
1.39
8.13
-0.13
-16.64
-0.09
-0.44
-1.87
-1.41
-0.02
-0.21
-0.01
-0.35
-1.38
-0.23
Error/Actual value
(if actual > 0)
-0.74
0.26
0.25
0.22
0.19
0.03
0.45
0.70
-0.94
-0.02
0.30
0.08
0.25
-0.48
0.24
0.34
-0.40
-0.15
0.23
0.37
-0.54
-0.23
Madagascar
Malaysia
Malta
Mauritius
Mexico
Mongolia
Morocco
Mozambique
Myanmar
Netherlands
New Zealand
Nicaragua
Norway
Pakistan
Panama
Paraguay
Peru
Philippines
Poland
Portugal
Romania
Russian Federation
Singapore
Slovak Republic
Slovenia
South Africa
Spain
St. Lucia
St. Vincent and the Grenadines
Sudan
Sweden
Switzerland
Tajikistan
Thailand
Trinidad and Tobago
Tunisia
Turkey
Uganda
Ukraine
Uruguay
Zambia
United Kingdom
United States
0
0
0
0
1
0
0
0
0
12
5
0
4
0
0
0
0
0
2
2
0
2
2
0
1
3
9
0
0
0
11
8
0
0
0
0
1
0
0
0
0
42
159
0.03
1.45
0.22
0.15
4.32
0.02
0.28
0.01
0.04
8.51
1.94
0.03
4.75
0.58
0.20
0.09
0.55
0.58
2.54
2.07
0.80
5.72
5.42
0.68
0.74
2.20
9.25
0.03
0.01
0.13
8.02
6.28
0.02
1.19
0.29
0.29
2.54
0.04
0.73
0.21
0.03
26.76
176.87
32
-0.03
-1.45
-0.22
-0.15
-3.32
-0.02
-0.28
-0.01
-0.04
3.49
3.06
-0.03
-0.75
-0.58
-0.20
-0.09
-0.55
-0.58
-0.54
-0.07
-0.80
-3.72
-3.42
-0.68
0.26
0.80
-0.25
-0.03
-0.01
-0.13
2.98
1.72
-0.02
-1.19
-0.29
-0.29
-1.54
-0.04
-0.73
-0.21
-0.03
15.24
-17.87
-3.32
0.29
0.61
-0.19
-0.27
-0.03
-1.86
-1.71
0.26
0.27
-0.03
0.27
0.22
-1.54
0.36
-0.11
Table 5 Estimation results of Tobit and Poisson models for TOP500
Tobit Model (f)
Poisson Model (f)
-26.677***
(2.072)
Model (f)
without US
-29.926***
(2.587)
Model (f)
without US and UK
-29.241***
(2.643)
-550.182***
(170.698)
LPOP
15.396***
(4.816)
0.766***
(0.060)
0.868***
(0.084)
0.848***
(0.086)
LGDPPC
29.232***
(8.914)
1.499***
(0.231)
1.640***
(0.235)
1.606***
(0.240)
R_D
4.477***
(1.617)
0.279***
(0.098)
0.292**
(0.123)
0.299**
(0.117)
ENG
15.022**
(7.396)
0.344*
(0.180)
0.575***
(0.138)
0.460***
(0.135)
2
0.602
0.768
0.726
0.730
0.743
0.965
0.839
0.843
93
93
92
91
Constant
R
~
R2
No. obs.
Same as Table 2.
Table 6 Estimation results for TOP100 and TOP300
Model (f)
TOP500
-26.677***
(2.072)
Model (f)
TOP300
-28.671***
(2.754)
Model (f)
TOP100
-36.212***
(5.451)
Model (f) TOP100
without US
-42.210***
(7.996)
Model (f) TOPp100
without US and UK
-41.883***
(9.576)
LPOP
0.766***
(0.060)
0.706***
(0.065)
0.666***
(0.087)
0.770***
(0.152)
0.674***
(0.132)
LGDPPC
1.499***
(0.231)
1.710***
(0.281)
2.271***
(0.521)
2.658***
(0.696)
2.789***
(0.838)
R_D
0.279***
(0.098)
0.377***
(0.113)
0.807***
(0.148)
0.888***
(0.179)
0.886***
(0.168)
ENG
0.344*
(0.180)
0.768
0.669***
(0.248)
0.738
1.091***
(0.357)
0.753
1.308***
(0.417)
0.753
0.701*
(0.396)
0.761
0.965
0.957
0.976
0.976
0.976
93
93
93
92
91
Constant
2
R
~
R2
No. obs.
Same as Table 2.
33
United States
United Kingdom
Germany
Japan
France
Italy
Canada
China
Australia
Netherlands
Sweden
Spain
Korea, Rep.
Switzerland
Austria
Belgium
Brazil
Finland
Israel
Hong Kong, China
New Zealand
Denmark
Norway
Ireland
South Africa
Chile
Greece
Hungary
India
Poland
Portugal
Russian Federation
Singapore
Argentina
Czech Republic
Mexico
Slovenia
Turkey
Figure 1 The actual numbers of top 500 universities and the predicted errors for 38
countries
180
160
140
Actual ‐ Predicted TOP500
40
Actual TOP500
120
100
80
60
40
20
0
‐20
Figure 2 The actual numbers of top 100 universities and the predicted errors for 15
countries
60
50
Actual‐Predicted Top100
30
20
10
0
‐10
34
Actual Top100
Table A1 Data for 93 countries
Country
Top500
Top300
Top100
LPOP
LGDPPC
R_D
ENG
MIG
EDUGDP
LEXPTER
Argentina
1
1
0
17.44
9.18
0.43
0
4.09
4.21
7.25
TERPOP
5.13
Armenia
0
0
0
14.93
7.94
0.23
0
9.47
3.22
6.91
2.29
Australia
15
9
3
16.79
10.30
1.66
1
20.91
4.68
8.85
4.76
Austria
7
2
0
15.90
10.40
2.07
0
12.59
5.70
9.63
3.04
Azerbaijan
0
0
0
15.91
7.97
0.33
0
2.04
3.55
6.06
1.45
Belarus
0
0
0
16.11
8.77
0.69
0
12.64
5.87
7.47
4.53
Belgium
7
6
0
16.15
10.32
1.95
0
8.06
6.05
9.32
3.54
Bolivia
0
0
0
15.96
8.18
0.29
0
1.14
5.93
7.34
3.54
Brazil
6
2
0
19.00
8.99
0.91
0
0.38
3.91
8.24
1.89
Brunei Darussalam
0
0
0
12.75
10.78
0.02
0
31.82
5.27
NA
1.25
Bulgaria
0
0
0
15.89
8.93
0.51
0
1.28
3.45
7.35
3.07
Cambodia
0
0
0
16.39
7.01
0.05
0
1.95
1.60
6.18
0.25
Canada
21
18
4
17.26
10.40
1.98
1
18.31
5.46
9.66
4.02
Chile
2
0
0
16.57
9.29
0.60
0
1.23
3.94
7.51
3.27
China
18
6
0
20.97
8.02
1.01
0
0.04
1.91
7.92
0.79
Colombia
0
0
0
17.57
8.60
0.18
0
0.28
4.71
7.52
2.29
Costa Rica
0
0
0
15.21
9.02
0.36
0
8.60
4.83
8.29
1.90
Croatia
0
0
0
15.31
9.33
1.12
0
14.04
4.48
8.39
2.38
Cyprus
0
0
0
13.47
10.06
0.29
0
12.67
6.01
9.34
2.01
Czech Republic
1
1
0
16.14
9.77
1.21
0
4.42
4.21
8.62
2.66
Denmark
4
3
2
15.49
10.37
2.43
0
6.14
8.33
10.00
3.69
Ecuador
0
0
0
16.34
8.67
0.06
0
0.83
1.37
NA
NA
Egypt, Arab Rep.
0
0
0
18.04
8.36
0.19
0
0.25
4.81
NA
3.44
4.29
Estonia
0
0
0
14.12
9.43
0.73
0
17.25
5.63
8.02
Finland
6
1
1
15.46
10.24
3.34
0
2.71
6.28
9.28
5.41
France
23
14
3
17.90
10.29
2.17
0
10.65
5.71
9.11
3.46
3.12
Georgia
0
0
0
15.35
7.88
0.24
0
4.52
2.28
NA
Germany
40
24
6
18.23
10.30
2.47
0
12.04
4.57
NA
NA
Greece
2
1
0
16.21
10.15
0.50
0
7.32
3.85
8.80
4.52
Hong Kong, China
5
3
0
15.72
10.32
0.58
1
41.57
4.21
9.92
2.17
Hungary
2
0
0
16.14
9.59
0.86
0
2.97
5.16
8.42
3.42
Iceland
0
0
0
12.56
10.35
2.75
0
6.22
6.99
9.14
3.96
35
India
2
0
0
20.76
7.51
0.73
1
0.59
4.08
7.20
Indonesia
0
0
0
19.16
7.95
0.06
0
0.13
2.77
NA
1.01
1.54
Iran, Islamic Rep.
0
0
0
18.00
9.01
0.59
0
3.39
4.66
7.88
2.42
Ireland
3
1
0
15.18
10.43
1.16
1
11.30
4.36
9.12
4.37
Israel
6
4
1
15.69
9.99
4.42
0
36.64
7.21
8.77
4.30
Italy
22
7
0
17.86
10.23
1.08
0
3.30
4.66
8.85
3.24
Jamaica
0
0
0
14.77
8.64
0.06
1
0.70
5.40
8.16
1.62
Japan
31
12
4
18.66
10.27
3.12
0
1.37
3.66
8.56
3.13
Jordan
0
0
0
15.42
8.24
0.34
0
40.70
4.94
NA
3.46
Kazakhstan
0
0
0
16.52
8.77
0.22
0
18.46
3.09
6.32
3.27
Korea, Rep.
8
3
0
17.67
9.83
2.54
0
1.19
4.31
7.29
6.34
Kuwait
0
0
0
14.65
10.52
0.17
0
64.11
6.30
10.82
1.54
Kyrgyz Republic
0
0
0
15.42
7.36
0.19
0
6.98
3.67
5.75
3.67
Latvia
0
0
0
14.67
9.19
0.40
0
21.79
5.45
7.55
4.50
Lesotho
0
0
0
14.47
7.11
0.05
1
0.29
11.03
9.26
0.26
Lithuania
0
0
0
15.06
9.28
0.64
0
5.69
5.53
7.97
4.15
Luxembourg
0
0
0
13.00
11.07
1.66
0
37.08
3.35
NA
0.63
Macao, China
0
0
0
13.02
10.15
0.07
0
54.46
3.21
9.22
3.67
Madagascar
0
0
0
16.64
6.72
0.20
1
0.37
2.95
7.37
0.20
Malaysia
0
0
0
16.99
9.26
0.59
0
6.10
7.01
9.18
2.54
Malta
0
0
0
12.89
9.90
0.35
1
2.35
4.78
9.06
1.84
Mauritius
0
0
0
14.00
9.10
0.36
1
1.42
4.01
8.32
1.04
Mexico
1
1
0
18.42
9.29
0.42
0
0.58
5.15
8.46
2.10
Mongolia
0
0
0
14.71
7.68
0.25
0
0.35
7.04
6.48
3.55
Morocco
0
0
0
17.18
8.07
0.60
0
0.42
6.32
8.07
1.06
Mozambique
0
0
0
16.75
6.34
0.50
0
2.00
3.33
8.01
0.08
Myanmar
0
0
0
17.66
6.65
0.10
0
0.25
0.82
5.39
1.08
Netherlands
12
9
2
16.59
10.42
1.81
0
9.88
4.90
9.51
3.16
New Zealand
5
2
0
15.19
10.05
1.11
1
17.50
6.73
9.06
4.82
Nicaragua
0
0
0
15.47
7.63
0.05
0
0.53
3.55
NA
1.91
Norway
4
2
1
15.33
10.71
1.64
0
6.89
7.27
9.94
4.39
Pakistan
0
0
0
18.78
7.58
0.16
1
2.78
2.10
NA
0.29
Panama
0
0
0
14.93
9.04
0.33
0
3.00
4.48
7.86
3.96
Paraguay
0
0
0
15.52
8.23
0.09
0
3.15
4.83
7.28
2.06
Peru
0
0
0
17.08
8.65
0.11
0
0.17
2.98
6.74
3.20
36
Philippines
0
0
0
18.18
7.90
0.14
1
0.43
3.16
5.93
3.02
Poland
2
0
0
17.46
9.39
0.60
0
2.05
5.20
7.77
4.65
Portugal
2
0
0
16.15
9.89
0.76
0
6.52
5.47
8.56
3.73
Romania
0
0
0
16.91
8.94
0.39
0
0.60
3.32
7.60
2.51
5.79
Russian Federation
2
1
1
18.79
9.15
1.15
0
8.22
3.42
6.98
Singapore
2
1
0
15.23
10.54
2.07
1
36.45
3.67
NA
NA
Slovak Republic
0
0
0
15.50
9.51
0.60
0
2.23
4.15
8.33
2.72
Slovenia
1
0
0
14.50
9.92
1.45
0
8.65
6.01
8.61
4.68
South Africa
3
1
0
17.62
8.95
0.80
1
2.33
5.42
8.30
1.51
Spain
9
3
0
17.53
10.15
0.96
0
6.14
4.28
8.61
4.44
St. Lucia
0
0
0
11.97
8.98
0.41
1
4.90
6.67
NA
1.34
St. Vincent and the
Grenadines
0
0
0
11.67
8.63
0.10
1
7.16
9.55
NA
NA
Sudan
0
0
0
17.35
7.29
0.41
1
2.31
NA
NA
0.61
Sweden
11
9
4
16.00
10.29
3.88
0
11.55
7.39
9.58
4.24
Switzerland
8
7
3
15.80
10.44
2.75
0
21.92
5.66
9.91
2.36
Tajikistan
0
0
0
15.65
7.05
0.07
0
5.15
2.46
5.17
1.39
Thailand
0
0
0
17.93
8.71
0.26
0
1.47
4.92
7.51
3.37
Trinidad and Tobago
0
0
0
14.08
9.53
0.12
1
3.08
3.97
9.66
0.83
Tunisia
0
0
0
16.09
8.65
0.63
0
0.39
6.99
8.35
2.27
Turkey
1
0
0
18.05
9.08
0.66
0
1.86
3.74
8.31
2.47
Uganda
0
0
0
17.07
6.66
0.32
1
2.04
3.85
7.29
0.25
Ukraine
0
0
0
17.70
8.36
1.02
0
14.24
4.80
7.36
4.27
Uruguay
0
0
0
15.01
9.04
0.25
0
2.64
2.68
7.42
2.97
Zambia
0
0
0
16.19
6.98
0.01
1
3.11
2.18
7.50
0.23
United Kingdom
42
33
11
17.90
10.28
1.82
1
8.36
5.02
8.95
3.64
United States
159
114
54
19.47
10.58
2.68
1
12.51
5.60
9.26
5.23
37